New MOSAIC architecture improves multi-task learning by separating modality dependencies

Researchers have developed MOSAIC (MOdality-Separated Architecture with Inter-modal Cross-attention), a novel approach to multi-task learning that explicitly models task-modality dependencies. Unlike traditional fusion-based architectures that enforce a unified representation space, MOSAIC preserves unimodal representations while allowing controlled interaction. This method uses modality-specific losses to route gradients to their corresponding partitions, preventing interference between tasks with varying reliance on different modalities. In experiments on the Corpus of Everyday Japanese Conversation, MOSAIC achieved a 7.4-point improvement in average macro-F1 across four tasks, with a 72% reduction in backbone parameters compared to a fusion-based baseline. AI

IMPACT This architecture could lead to more efficient and effective multi-modal AI systems by better handling tasks with varying modality requirements.

RANK_REASON Research paper detailing a new model architecture and its performance improvements. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• alphaXiv
• CatalyzeX Code Finder for Papers
• CORE Recommender
• Corpus of Everyday Japanese Conversation
• DagsHub
• Gotit.pub
• Hugging Face
• IArxiv Recommender
• Influence Flower
• MOSAIC
• ScienceCast

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →